Luminance normalization?

In some training and validation pairs, we can observe that there are very noticeable differences between the HE and IHC images in terms of brightness.

For example: 00788_train_3+.png and 00746_train_1+.png

Will cases like this appear in the test set as well? This discrepancy in brightness will inadvertently affect both PSNR and SSIM.

Yes, there are several such image pairs in the test set too. This is a common phenomenon that occurs during pathological image scanning, therefore, we use the average PSNR and SSIM to measure the overall quality of the generated image, reducing the impact of brightness inconsistencies in individual image pairs.